Could giving people an AI's track record or blind spots help them judge when to trust its predictions?
Can contextual information help humans integrate AI predictions better?
This explores whether giving people extra context alongside an AI's prediction (such as its track record, what it doesn't know, or the situation around the prediction) helps them decide when to trust it and how to combine it with their own judgment.
This explores whether context around an AI prediction helps people use it well, rather than over-trusting or ignoring it. The retrieved material has no study that directly tests humans weighing AI predictions with and without added context, so it can't answer this head-on. What it does show is which kinds of context make a prediction easier to judge in the first place. A person can only weigh an AI's prediction well if that context exists and comes with the prediction.
The most useful kind of context is a track record. Can past performance predict when a model will be right? finds that a model's confidence in a single answer is a weak signal. A much better one comes from looking up how often the model was right in past cases where it felt similarly sure. The ablations show the gain comes entirely from those stored outcomes, not from clever prompting. Applied to human-AI teams, the most useful thing to show a person next to a prediction may not be an explanation. It may be a plain record like "in situations like this, it was right 6 times out of 10."
A second kind of context is being clear about what the AI doesn't know. Do language models know what they don't know about users? shows that assistants hallucinate and flatter users partly because they have no explicit list of unknowns. Adding a schema of labeled unknowns cut harmful advice by 50–75%. Those same labels would tell a human reader exactly where a prediction is guessing, which is where their own knowledge should take over.
A third idea is to keep the numbers and the story separate. Can decomposing forecasting into stages unlock numerical and contextual reasoning? gets better forecasts by not making one model handle both trend-extrapolation and event-driven reasoning. It splits them into stages and combines them at the end. That suggests a design for human-AI work: let the AI supply the numerical prediction, let the human supply the situational context they uniquely have, and make the step where the two are combined explicit.
There's a twist worth knowing. Integrating context is hard for models too. Why do language models ignore information in their context? shows that when a model's training-based habits are strong, it ignores the facts right in front of it, and prompting alone won't fix that. The social-norms work gives the human-side mirror image. GPT-4.5 judges social appropriateness better than any individual person, yet every model makes the same systematic mistakes on unwritten norms (Can AI learn social norms better than humans?, Can AI predict social norms better than humans?). The best context to give a person, then, may be a map of where the AI is reliably blind, because those are the places where their own judgment is worth the most.
Sources 6 notes
XConf matches ten-sample self-consistency at a tenth of the cost by retrieving the model's past episodes with similar confidence levels and reading their historical success rates. Ablations show the signal depends entirely on stored outcomes, not on the retrieval prompt itself.
Research shows assistants suffer from sycophancy and hallucination because they have no representation of what remains unknown about users. Adding a schema of labeled unknowns to prompts reduced harmful advice and sycophancy by 50–75% and cut hallucination rates by roughly half.
Nexus outperforms pure TSFM and LLM baselines on real-world datasets by decomposing forecasting into contextualization, dual-resolution macro/micro outlook, and synthesis stages. Separating numerical extrapolation from event-driven contextual reasoning avoids forcing one model to handle both simultaneously.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
GPT-4.5 outperformed every individual human at judging social appropriateness across 555 scenarios, challenging the theory that embodied cultural experience is necessary. However, all AI models share identical systematic errors on unwritten norms.
Show all 6 sources
GPT-4.5 outperforms all individual humans at predicting social appropriateness, yet structurally cannot enter the community processes that establish and validate norms. This reveals a critical gap between pattern-matching and authentic participation in knowledge-making.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms
- Humans learn to prefer trustworthy AI over human partners
- SOTOPIA: Interactive Evaluation for Social Intelligence in Language Agents
- Harnessing the power of LLMs for normative reasoning in MASs
- Norm Enforcement for AI Agents: Robustly Shaping Behavior in Multi-Agent Systems
- Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents
- Nexus: An Agentic Framework for Time Series Forecasting
- Reported Confidence in LLMs Tracks Commitment More Than Correctness